We employ roofline‑shaped predictors to estimate single‑sequence llama.cpp throughput directly from GGUF metadata. The predictors incorporate quantization‑specific scale factors fitted on reference models. Our dataset comprises 318 phase‑depth measurements across 53 host‑file configurations collected on two Apple M4 Max machines and an NVIDIA RTX 5080 GPU.
On host‑specific held‑out sets of four, five, and two configurations, an active‑parameter decode model achieves mean absolute percentage error (MAPE) of 13.1%, 14.4% and 36.1%, compared with 49.4%, 55.3% and 51.9% when using total parameter counts alone.
Leave‑one‑host‑out coefficients fitted on the other two systems yield test MAPE of 11.6%, 16.8% and 36.0%. A low‑bit model ladder changes ordering across runtime stacks. The P2 prefill baseline records test MAPE of 18.7%, 22.2% and 108.2%.
GGUF structural information benefits all three systems, yet the fitted efficiency factors are not universal.
$$\text{MAPE}=\frac{1}{n}\sum_{i=1}^{n}\left|\frac{y_i-\hat{y}_i}{y_i}\right|\times100\%$$
Review